Developing With TensorFlow

This repository tackles a variety of projects in order to get familiar with the TensorFlow Framework.

1. Regression With TensorFlow

This section introduces the basics of tensorflow and many methods that are incorporated in the framework from other modules. Later on, a simple introduction the steps to building a model using the Sequential Model approach. The main takeaways are:

• Creating a Model, compiling it and fitting it
• Learning how to expand the dimensions or squeeze it
• Predicting and Evaluating the model
• Understanding different optimizers and metrics

2. Classification With TensorFlow

This section further elaborates on the Sequential Modelling in TensorFlow. The key takeaways are:

• Tweak the model after evaluating it
• Visualize data to further understand it
• Finding the best learning rate (around 10x before the lowest point)
• Building Binary and Multi-Class Classification Models

3. Computer Vision With TensorFlow

This section focuses preprocessing images into batches and building a CNN model: The main points are:

• Using ImageDataGenerator from tensorflow to rescale images and batchify them
• Applying Data Augmentation to images
• Building binary and multi-class classification models
• Trying out custom data on the model

4. Transfer Learning With TensorFlow

This section focuses on best practices when building a model for doing tasks:

• Setting up callbacks
• Loading pretrained models using Tensor Hub
• Testing out ResNet vs Efficientnet
• Using less validation data during fitting to speed up the process

5. Transfer Learning With TensorFlow Part 2

This section attempts to compare the results of a pre-trained model by adding new parameters:

• Trying varying amounts of train data
• Implementing data augmentation
• Fine-Tuning the loaded pre-trained model
• Utulizing the full data

6. Transfer Learning With TensorFlow Part 3

This section attempts to tackle a much bigger dataset (10x the previous section) with different models:

• Trying varying amounts of train data
• Visualizing each step of the process with different visual techniques
• Finding the most wrong predictions (model-driven data exploration)

7. Transfer Learning Capstone

This section attempts to build a baseline for you to improve on:

• Building a pre-trained baseline model for the Kaggle 101 Food dataset (70% accuracy)
• Fine-Tune the built model to beat the baseline (73% accuracy)
• Try improving the model even further? (Add dropout layers, further augmentation, longer training time, unfreeze more layers)
• Load and test model made in the notebook for less clutter. How about custom data?
• Try different architectures e.g EfficientNetv2 (try to avoid the larger models like sticking to B0/lite versions)
• After even further improvements 75% accuracy, EfficientNetv2 77% accuracy

8. NLP with TensorFlow

This section introduces the field of NLP and its implementation in TensorFlow:

• Creating a tokenization vector (mapping words to numbers)
• Building an embedding layer (creating a feature vector matrix for each token)
• Utulizing bidirectional layers
• Transfer Learning with pre-trained models